Evaluation of soil quality with artificial neural networks thecase of Bingöl Solhan Plain

In this study, in order to model some soil properties in Bingöl Solhan Plain, soil samples at 0-30 cm depth were taken from 85 points using 300 m x 300 m spaced grids with ArcGIS 10.8 software and the general soil quality of the plain was modeled with soil quality parameters determined by expert opinion. Physical, chemical and biological analyses of soil samples were carried out and each parameter was normalized and soil quality index was calculated using SMAF model and additive index method. According to the results; physical quality (average score of 60) was classified as "medium", chemical (score of 54) and biological (score of 48) quality as 'low' and overall soil quality (score of 56) as "medium". Improvement of physical, chemical and biological quality is necessary for the functionality of soils. This depends on appropriate land management and management systems. Poor practices can further degrade already poor properties. The study found that the SMAF model is an effective tool for soil quality assessment and facilitates sustainable decision making for farmers/land managers. In the second stage of the research, soil properties were estimated by artificial neural networks (ANN). Levenberg-Marquardt (LM), Scaled Conjugate Gradient (SCG) and Bayesian Regularization (BR) algorithms were used to optimize the hyperparameters. Double hidden layer networks with 0-30 neurons were tested with 70% training, 15% testing and 15% validation data. The LM algorithm showed the best performance in all predictions: Percent exchangeable cation (85% accuracy, 1.705 MAE, 2.255 RMSE), cation exchange capacity (73% accuracy, 0.439 MAE, 0.533 RMSE), field capacity (97% accuracy, 0.560 MAE, 1.012 RMSE) and wilting point (99% accuracy, 0.003 MAE, 0.004 RMSE). LM provided the highest R² and lowest error values in all models. Thus, it was proved that costly and specialized soil parameters can be estimated with high accuracy.